An ASIC — Application-Specific Integrated Circuit — is a chip designed for one specific job and nothing else. In video, ASICs are custom silicon built to transcode video at massive scale, far more efficiently than general-purpose CPUs or GPUs. They sacrifice flexibility (you can't run anything else on them) for breathtaking performance per watt and per dollar. At hyperscale, that's the trade-off video infrastructure teams happily make.
The famous example is YouTube's Argos, a transcoding ASIC built by Google specifically for processing the unimaginable volume of video YouTube ingests every minute (over 500 hours per minute as of 2023). A single Argos chip transcodes video at speeds that would take racks of CPUs, at a fraction of the power. Meta has similar internal ASICs. Commercial offerings exist too — NETINT's Quadra cards combine ASIC video silicon with general-purpose I/O so any cloud transcoding pipeline can plug them in.
For a product team, video ASICs are the answer when you've outgrown both CPU and GPU economics. Below maybe 10,000 hours of video per day, software encoders on CPU or hardware encoders on GPU make more sense — flexible, available, cheap to start. Above that scale, the per-watt and per-rack-unit savings of dedicated ASIC silicon start dominating: a NETINT Quadra card might transcode 80 concurrent 1080p streams at 50W, where the equivalent CPU work would need a full server at 400W. For most companies this is purely an "if your traffic explodes, look here" consideration; for the hyperscale platforms, it's already standard.
